Model
Model explorer

MiniMax M3.5

OPEN
MiniMax · MiniMax M3 family · released Sep 8, 2026

MiniMax's September 8, 2026 step up from June's M3 — a modest scale-up to 458B/25B (from 428B/23B) under the same non-commercial community license, with the API price moving 0.30/1.20 → 0.35/1.40. The launch table keeps MiniMax's telltale mix: big self-reported gains on the agentic/video rows it cares about (Video-MME, BrowseComp, OSWorld-Verified) plus the now-standard trio of independent rows (GPQA, HLE, IFBench) that land where expected for a 25B-active model. No architecture change; the pitch is 'frontier-adjacent agentic quality at non-commercial-license prices.'

ReasoningCodingVisionFunction callingTool useAgentic
2544.7
Elo · rank #31
Parameters
458B
Active params
25B (MoE)
Context
1.024M tokens
Architecture
458B-total/25B-active sparse MoE, open weights under the MiniMax community license; 1.02M-token context, natively video-capable
License
MiniMax Model Community License (non-commercial redistribution terms apply)
Languages
—
API price (in/out)
$0.35 / $1.4
Modalities
text · vision · video
Benchmark results
Bar shows position within the tracked field; marker = field best
BrowseCompAgents84.9%#11
best: Kimi K3.1 · 92.4%
GPQA DiamondReasoning93.4%#12
best: GPT-6 Astra · 96.0%
Humanity's Last ExamReasoning39.2%#40
best: Claude Opus 5 · 64.7%
IFBenchReasoning84.1%#2
best: Inkling-Medium · 85.3%
MMMU-ProVision79.8%#16
best: Claude Opus 4.7 · 85.5%
OSWorld-VerifiedAgents77.4%#10
best: Claude Fable 5 · 85.0%
SWE-bench ProCoding61.4%#19
best: Claude Fable 5.1 · 81.2%
SWE-bench VerifiedCoding81.9%#11
best: Claude Opus 5 · 96.0%
Terminal-Bench 2.0Coding69.3%#35
best: Gemini 3.8 Flash · 89.4%
USAMOMath87.2%#2
best: Claude Opus 4.8 · 96.7%
Video-MMEVision86.7%#7
best: Seed 2.1 Pro · 89.2%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
Research only
QLoRA311.4 GB2× B200 192GB
LoRA975.5 GB8× H200 141GB
Full fine-tune7350.9 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $8.94 (2× B200 192GB)
Plan a fine-tune →
MiniMax M3 family
Elo progression across releases
MiniMax M3Jun 20262435.1MiniMax M3.5Sep 20262544.7
Resources
HFMiniMaxAI/MiniMax-M3.5DOCShttps://platform.minimax.io/docs/modelsAPIAvailable via MiniMax API
API price $0.35/$1.4 · each benchmark row carries its own source badge (see methodology)